Overview
Ornith-1.5-9B is the smallest model in the Ornith 1.5 family, released on August 18, 2026 under MIT. Ornith 1.0 was built on top of Qwen3.5 and Gemma 4 with extra pretraining and post-training; 1.5 extends it with a self-improvement loop where the model proposes its own training tasks and learns from its own attempts.
| Specification | Ornith-1.5-9B |
|---|---|
| Total parameters | 9.4B |
| Architecture | Dense, hybrid attention, built on the Qwen3.5 stack |
| Layers | 32 |
| Context window | 262,144 tokens |
| Modalities | Text and image input |
| Reasoning | Thinking trace returned separately from the answer |
| Release date | August 18, 2026 |
| License | MIT |
What that training approach buys is agentic behaviour at a size that fits a laptop. The model returns its thinking trace in a separate field from the answer, so a harness can log the reasoning without pasting it into the chat.
Ornith-1.5-9B benchmarks
Ornith's launch numbers, from the model card, put the 9B against the model it replaces, the Qwen3.5 9B it was built from, and two models three to four times its size:
| Ornith-1.5-9B | Ornith-1.0-9B | Qwen3.5-9B | Qwen3.6-35B-A3B | Gemma-4-31B | |
|---|---|---|---|---|---|
Terminal-Bench 2.1 Agentic terminal-use tasks | 46.2 | 43.1 | 21.3 | 52.5 | 42.1 |
SWE-bench Verified Real-world software engineering tasks | 70.6 | 69.4 | 53.2 | 73.4 | 52.0 |
SWE-bench Pro Real-world software engineering, harder split | 47.5 | 42.9 | 31.3 | 49.5 | 35.7 |
GPQA Diamond Expert-level science questions | 86.4 | 82.5 | 81.7 | 86.0 | 84.3 |
Humanity's Last Exam Expert-level questions, tools allowed | 30.5 | 26.4 | 24.5 | 28.9 | 26.5 |
MCP-Atlas Tool use over MCP servers | 54.2 | 49.4 | 46.8 | 62.8 | 55.0 |
ClawEval Agentic coding in a real harness | 66.5 | 63.1 | 53.2 | 68.7 | 48.5 |
The 9B improves on Ornith-1.0-9B on every row and clears Qwen3.5-9B by a wide margin. Against models three to four times its size it still takes GPQA Diamond and the tool-assisted HLE split, and gives up ground on the agentic rows.
Ornith-1.5-9B hardware requirements
The system requirement to check is memory. We quantized the model from the original weights and published the builds as AtomicChat/Ornith-1.5-9B-GGUF.
| Memory | Build to pick | File size |
|---|---|---|
| 8 GB | AD-IQ4_XS | 5.5 GB |
| 12 GB | Q5_K_M | 6.5 GB |
| 16 GB | Q6_K | 7.4 GB |
| 24 GB and up | Q8_0 | 9.5 GB |
At this size the whole ladder fits on ordinary hardware, so there is little reason to go below AD-IQ4_XS. A 16 GB MacBook runs Q6_K with room for a long chat; see the best local LLMs for a 16 GB Mac for what else fits alongside it.
How to run Ornith-1.5-9B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Ornith-1.5-9B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
The full walkthrough, with the quant table and the llama.cpp commands, is in our guide to running Ornith 1.5 9B locally.
The larger sibling is Ornith-1.5-35B-A3B, and both are listed on the Ornith family page. For how this class of model compares on code, see the best local LLMs for coding.
License
Ornith-1.5-9B is released under the MIT license. It permits commercial use, modification, redistribution and private deployment as long as the copyright notice stays with the code, which makes it one of the most permissive licenses a model can ship under.